CA2475189C

CA2475189C - Method and apparatus for window matching in delta compressors

Info

Publication number: CA2475189C
Application number: CA002475189A
Authority: CA
Inventors: Kiem-Phong Vo; Binh Dao Vo
Original assignee: AT&T Corp
Current assignee: AT&T Corp
Priority date: 2003-07-17
Filing date: 2004-07-19
Publication date: 2009-10-06
Anticipated expiration: 2024-07-19
Also published as: CA2475186A1; CA2475186C; CA2475189A1; US20110173167A1; US20050055367A1; US20050044294A1; US7925639B2; US8200680B2; CA2686618A1; US20080040375A1; US7296030B2; US7454431B2

Abstract

The present invention provides efficient target file window matching within the source file for delta compressors. One embodiment provides an apparatus for matching target file windows of a target file with source file windows of a source file in delta compressors, comprising: means for partitioning the source file into a plurality of source file windows; means for partitioning the target file into a plurality of target file windows; and means for matching one of said plurality of source file windows with one of said target file windows based upon a distance function that defines a distance between a data segment of said one of said plurality of source file windows and a data segment of said one of said target file windows using a respective n-gram of said data segments.

Description

METHOD AND APPARATUS FOR WINDOW MATCHING IN DELTA
COMPRESSORS
BACKGROUND OF THE INVENTION
Technical Field The present invention relates generally to data compression and, more particularly, to a method for efficient window partition matching in delta compressors to enhance compression performance based on the idea of modeling a dataset with the frequencies of its n-grams.
Description of the Problem Compression programs routinely limit the data to be compressed together in segments called windows. The process of doing this is called windowing. Delta compression techniques were developed to compress a target file given some other related source file. An early example delta compressor was the UNIX diff tool which computes differences between versions of text files by matching text lines across files. Diff was used to reduce disk storage in UNIX based source code control systems such as Source Code Control System (SCCS) and Revision Control System (RCS).
One of the first delta compressors that could deal with binary data was Vdelta.
Unlike other delta compressors which do nothing without source data, the algorithm used in Vdelta compresses data when source data is absent.
Vdelta was used in the Bell Labs source and binary code control system (SBCS) and later in the experiments that spurred the addition of delta-encoding to the Internet HTTP protocol. Windowing is necessary to delta compressors partly due to their use of memory-intensive string matching algorithms and partly due to improved coding of addresses of matches. Here, windowing means first dividing a target file into data segments called target windows, then for each target window selecting some suitably similar source window to compress against. Current practice uses source data with matching file offsets to given target windows. This may work for file versions created by simple editing but often fails when significant data updates happen between versions.

Therefore, a need exists for a method for efficiently matching target file windows with source file windows to improve delta compression performance.
SUMMARY OF THE INVENTION
In one embodiment, the present invention significantly improves the performance of matching target file window partitions to source file window partitions in delta compression. In particular, the present invention enhances delta compression performance of delta compressors based on the idea of modeling a dataset with the frequencies of its n-grams and employs a method to compute good source and target io window matching in time O(s).
Certain exemplary embodiments may provide a method for matching target file windows of a target file with source file windows of a source file in delta compressors, said method comprising: partitioning the source file into a plurality of source file windows; partitioning the target file into a plurality of target file windows;
and matching one of said plurality of source file windows with one of said target file windows based upon a distance function that defines a distance between a data segment of said one of said plurality of source file windows and a data segment of said one of said target file windows using a respective n-gram of said data segments.
Certain other exemplary embodiments may provide a computer-readable medium having stored thereon a plurality of instructions, the plurality of instructions including instructions which, when executed by a processor, cause the processor to perform the steps of a method for matching target file windows of a target file with source file windows of a source file in delta compressors, comprising of:
partitioning the source file into a plurality of source file windows; partitioning the target file into a plurality of target file windows; and matching one of said plurality of source file windows with one of said target file windows based upon a distance function that defines a distance between a data segment of said one of said plurality of source file windows and a data segment of said one of said target file windows using a respective n-gram of said data segments.

-2a-Yet another exemplary embodiment may provide an apparatus for matching target file windows of a target file with source file windows of a source file in delta compressors, comprising: means for partitioning the source file into a plurality of source file windows; means for partitioning the target file into a plurality of target file windows; and means for matching one of said plurality of source file windows with one of said target file windows based upon a distance function that defines a distance between a data segment of said one of said plurality of source file windows and a data segment of said one of said target file windows using a respective n-gram of said data segments.
BRIEF DESCRIPTION OF THE DRAWINGS
The teaching of the present invention can be readily understood by considering the following detailed description in conjunction with the accompanying drawings, in which:
FIG. 1 illustrates an example of window matching using the file offset alignment strategy;
FIG. 2 illustrates a flowchart of a fast search for source window matching method of the present invention;
FIG. 3 illustrates an exemplary partition of the source and target files of the present invention;
FIG. 4 illustrates the candidate source window matches with the target window T and the associated votes by target each target segment in target window T;
FIG. 5 illustrates a detailed window matching algorithm in C programming codes of the present invention;
FIG. 6 illustrates a flowchart of a best match method for finding the best match between the target segment T among the top source window candidates;
FIG. 7 illustrates an example of the extended search area for the top source window candidates and the associated target window T of the present invention;
and FIG. 8 illustrates the present window matching method implemented using a general purpose computer or any other hardware equivalents.

To facilitate understanding, identical reference numerals have been used, where possible, to designate identical elements that are coimmon to the figures.
DETAILED DESCRIPTION
The present invention relates to data compression using delta compressors based on the idea of modeling a dataset with the frequencies of its n-grams.
To better understand the present invention, a description of n-grams and its use are first provided. The present invention uses n-grams to model data. For any data segment S with length s, an n-gram in S is a subsequence of n < s consecutive bytes. Assume an arbitrary but fixed n, the riotation S;
will denote the n-gram in S starting at location i while S[i] denotes the byte at location i. For example, the string S = abababac is of length 8 and has five 4-grams, of which three are distinct: abab, baba and abac. The 4-girams So and S2 are the same: abab.
The present invention repeatedly examine n-gram frequencies of given data segments. Thus, it is beneficial if this step can be executed quickly.
For any general data segment S, the notation Fs shall be used to denote an associative array of frequencies indexed by the n-grams of S. Suppose that Fs was initialized to 0's, the below loop computes all n-gram frequencies:
for(i=O; i<=s-n; i+=1) Fs[Si]+=1;
This loop runs in time O(s) as long as the cost of indexing the array Fs can be bounded by some constant. This can be ensured by implementing Fs as a hash table indexed by the distinct n-grams. However, hash table look-up cost is significant and the frequency estimates do not always need to be exact. Thus, Fs is chosen to be implemented as a normal array of size A by hashing each n-gram S; to an integer via the below hash function with some preselected constant a:

Proprietary Information of AT&T

x(Si) -(an-1S[i] + a-2S[i] +... + S[i+n-1]) mod A (Equ. 1) The above loop then becomes:

for(i=0; i<=s-n; i+=1) Fs[x(Si)]+=1;
For nontrivial values of n, the loop can be further optimized by exploiting the linearity of the hash function x to compute x(S;+,) from x(S;) via:
x(S,+,) a _{a(x(S~) - "-'S[i]} + S[i+n]} mod A (Equ. 2) For assessment and identification of string similarity in delta compressors, larger values of n are required (n usually greater than 1). In that case, a small number of n-grams may collide and get mapped into the same indices in Fs. Henceforth, given a data segment S and a frequency array FS, it shall be assumed that FS is indexed by mapping the n-grams via the x function as described. Therefore, the notation Fs[S;] will mean Fs[x(S;)].
Delta compressors typically deal with two files at a time. A target file is compressed against some related source file. For large files, windowing is done by first dividing the target file into target windows, then compressing each such window against some source window. A source window is often from the source file but it could also come from some part of the target file earlier than the current target window. Since the target file is often a Iater version of the source file created via local editing operations, windowing is conventionally done by aligning a target window with a source window in the source file using the same file offset. However, it is increasingly common to have more extensive changes between versions, for example, in applications that compute patches between archives of source code or store versions of daily disk back-up archives. The file offset alignment strategy is often ineffective for processing these very large files because data may be moved Proprietary Information of AT&T

to remote locations between versions. FIG. 1 illustrates an example of window matching using the file offset alignment strategy. The source file S
contains three string based sections A, B, and C ordered consecutively in sequence. The target file is simply a version of the source file by moving section A to the end of the file via local editing. The target file contains the same string based sections as the source file but the sections are ordered in the sequence of B, C, and A instead. If target window 1, T, is used to match against source window 1, a match cannot be found. In fact, the contents in the T have shifted quite far away from the source window 1. In this case, the file offset alignment strategy performs very poorly in matching a target window with a source window. A brute force approach is to simply align a target window with every location in the source file and run the compression algorithm itself to find the best one. This would be unbearably slow.
To address this criticality, the present invention provides a method for matching source and target file windows by modeling data using n-grams.
The present invention gives a way to construct fast linear-time window matching algorithms.
In the present invention, the concept of signatures to represent a data segment is introduced. Given a data segment P of length k, defirie its signature 6(P) as the sum of the hash values of all n-grams using the same hash function shown in Equ. 1 and 2. If i represents the hashed value of n-grams and A is the size of the n-gram frequency array, then:

6(P) = Y, FP(i] (Equ. 3) 1=0 The constant A is chosen so that the total sum would not E:xceed the maximum allowed size in a computer word. In one embodiment of the present invention, k is typically 210 so A = 220 is used to ensure that each signature fits in a 32-bit integer. A nice property of 6( ) is that similar data segments get mapped to similar values. The signature distance between two segments P, and P2 is defined as:

S(P15 P2) = 6(P,) -6(P21 (Equ. 4) max(cs(P,), 6(P2)) Proprietary Information of AT&T

FIG. 2 illustrates a flowchart of a fast search for matching source window method 200 of the present invention. Method 200 starts in step 205 and proceeds to step 210.
In step 210, a source file S is partitioned into a sequence of segments, So, Si, ... , Sa, each of fixed size k in length as shown in 310 in FIG. 3.
Note that the starting offset of segment S; in the source file is ik. The last segment is omitted if it is smaller than )~.
In step 220, a target file is k;~ in length and is split into L target windows as shown in 320 in FIG. 3. Each target window T is further split into a sequence of segments To, T7, ... , Tb each of the same size k as shown in 320 in FIG. 3. Note that the target window size is arbitrary and can be chosen to suit a particular application or implementation. Therefore, the values of b and L are arbitrary as well, depending on the value of the target window size chosen. In one embodiment of the present invention as shown in 320 in FIG.
3, the window size is chosen to be 3k and the values of b and L are therefore 3 and k/%/3 respectively.
In step 230, for each segment Tj in a particular target window T, the set of all Si's such that j<_ i<_ a- b and 8(S;, T;) <_ E are computed. Then, the method asserts Tj`s vote for each such source window with startirig offset (i-j)k as a candidate for a match. Note that different segments in T may vote for the same position. To further illustrate this step, the example in FIG. 4 is used.
In FIG. 4, the current target window T is target window 3. For segment To in T, To will be used to compute the signatures in conjunction with source segments So, Sl, S2, ... , Sa-b to produce b(So, To), b(Sl, To), 8(S2, To), ... , and 8(Sa-b, To)=
Similarly, for segment T, in T, T, will be used to compute the signatures in conjunction with source segments Sl, S2, S3, ... , Sa-b to produce 6(Si, Ti), 5(S2, Ti), 8(S3, Ti), ... , and S(Sa-b, T1). Similarly, for segment T2 in T, T2 will be used to compute the signatures in conjunction with source segments S2, S2, S4, ... , Sa-b to produce 8(S2, T2), 8(S3) T2), 8(S4, T2), ... , and 8(Sa-b, T2).
Considering an example with signatures of all source and target segment pairs that meet the condition of 8(S;, T;) <F, are shown in TABLE 1.
Proprietary Information of AT&T

Source Target Segment Pairs Meeting Condition of S(S;, T;) <_ c So, To S1, To S3, To S1, T1 Sg, T1 S2, T2 S5, T2 S11, T2 In this example, the candidate source windows that potentially match the current target window T and their associated votes by each target segment are shown in 430 to 434 in FIG. 4. The eight source and target window pairs that meet the 8(S;, T;) <_ c condition in this example produces five candidate matching source windows for the current target window.
For target window segment To, it has a potential match wit:h So, S1, and S3. In this case, j=0 and i=0, 1, and 3. Applying the previously defined rule of voting, then the To and So segment pair leads to 1 vote by target window segment To for candidate source window 1 which contains source segments So, S1, and S2; the To and S, segment pair leads to 1 vote by target window segment To for candidate source window 2 which contains source segments S1, S2, and S3; the To and S3 segment pair leads to 1 vote by target window segment To for candidate source window 3 which contains source segments S3, S4, and S5.
For target window segment T1, it has a potential match with S1 and S.
In this case, j=1 and i=1 and 8. Applying the previously defined rule of voting, then the T1 and S1 segment pair leads to 1 vote by target window segment T1 for candidate source window 1 which contains source segments So, S1, and S2; the T1 and S8 segment pair leads to 1 vote by target window segment T1 Proprietary Information of AT&T

for candidate source window 4 which contains source segments S7, S8, and ~'Jg.

For target window segment T2, it has a potential match with S2, S5, and S11. In this case, j=2 and i=2, 5, and 11. Applying the previously defined rule of voting, then the T2 and S2 segment pair leads to 1 vote by target window segment T2 for candidate source window 1 which contains source segments So, S1, and S2; the T2 and S5 segment pair leads to 1 vote by target window segment T2 for candidate source window 3 which contains source segments S3, S4, and S5; the T2 and S11 segment pair leads to 1 vote by target window segment T2 for candidate source window 5 which contains source segments S9, S1o, and S11.
At the end of step 230, the method produces the voting results by the target window segments of the current target window and the associated candidate matching source windows as shown in TABLE 2.
Candidate Matching Source Window Segments Number of Votes by Current Source Windows Target Window 1 So,S1,S2 3 2 S1, S2, S3, 1 3 S3, S4, S5 2 4 S7, S8, S9 1 5 S9,S1o,s11 1 Note that votes are weighted by the number of agreeing neighbors to increase accuracy, such as the case of candidate source window 1. The results of this example mean that candidate source window 1 has the highest potential match with the current target window T with 3 votes. The second highest potential match is candidate source window 3; the third highest potential match is a tie consisting candidate source windows 2, 4õ and 5.
These candidate matching source windows are the top 5 match of the candidate source windows.

Proprietary Information of AT&T

Note also that large signatures often mean small values even if they represent fairly distinct data segments. Thus, a graduated scheme is used to set the threshold c so that larger signatures get smaller thresholds. For example, in comparing two signatures 61 and a2, one embodimerit of the present invention uses c = 0.05 when min(6l, 62) > 224 and c = 0.08 when min(6l, 62) < 216. Experimentation with a wide variety of data showed that these parameters provided the best balance between window selections and running time. However, these parameters can be arbitrarily chosen to suit a particular application or implementation.
In step 240, the top k candidate source window positions identified in step 230 that have the closest signature distance from a target wiindow T will be applied as inputs to the detailed matching algorithm to a small neighborhood of the candidate source windows to find the best miatch to T.
The detailed matching method is discussed in detail in FIG. 5 and 6. In one embodiment of the present invention, the number of selected top candidates k = 8 is used and 2k (i.e., twice the length of a segment) for the search size for a candidate's neighborhood that will be used in the detailed matching method to be discussed later. Experinientation with a wide variety of data showed that these parameters provided the best balance between window selections and running time. However, these parameters can be arbitrarily chosen to suit a particular application or implementation.
The key step in matching windows is to measure their similarity. This is done by defining a distance between two data segments using their n-grams.
Let S, and S2 be two non-ernpty data segments with corresponding n-gram frequency arrays FS1 and FS2 computed by hashing n-grams as described in Equ. 1 and 2 and related exemplary programming loops described previously. The distance between S, and S2 is defined as follows::

i:AO FS [i] - FSZ [i]
0(Si, S2) _ (Equ. 5) ~AO' max FSji], FSz [i]) 30 Proprietary Information of AT&T

The distance A is normalized so that 05 A(SI, S2) <_ 1. This enables using some preset threshold to cut short a search for similar data segments.
The parameters defining n-grams and frequency arrays can have large effect on the quality of A. In one embodiment of the present invention, ri = 4 is used since that is the minimum match length allowed by the underlying delta compressor. The parameters a and A were set to be 101 and 213 . These choices keep the frequency arrays at a reasonable size with few n-gram collisions. Although A can be applied to data segments with different lengths, the use of A by the present invention to be shown below will always be on segments with the same length.
Considering an example of two data segments S and T with sizes s> t and a sub-segment of S with length t that matches T best needs to be found.
The detailed matching algorithm in FIG. 5 shows how this can be done.
FIG. 5 illustrates a detailed window matching algorithm 500 in C
programming codes of the present invention. The first loop computes the distance between T and an iriitial segment W of S of length t. The variables d and m keep the values of the sums used in defining the distance A. Then, the second loop slides W over S one position at a time. At each step, W drops off the left-most n-gram and adds a new n-gram on the right. The frequency array FW and the variables d and m can be updated in constant time to reflect this.
Therefore, the entire loop runs in time O(s). Omitted in the algorithm was the O(t) cost of computing the initial frequency arrays. Adding this back in, the algorithm, when given two data segments S and T with sizes s > t, can find a best match to a target windovv T in S in time O(s+t).
FIG. 6 illustrates a flowchart of a best match method 600 using the top candidate source windows found by method 200 and the associated target window T as inputs to find the best match between the target segment T
among the top source window candidates. Method 600 starts in step 605 and proceeds to step 610.
In step 610, the method accepts the top candidate source windows found by method 200 and the associated target window T as inputs.
In step 620, the method extends the search neighborhood of each of the top candidate source windows by 2k to give extended search windows Proprietary Information of AT&T

S's. For instance, for each of the top k source window candidate found in method 200, the method uses 2k (i.e. twice the length of a segment) for the search size for a candidate's neighborhood. In other words, the method uses a candidate source window generated from method 200 and extend the search area k to the right and k to the left of the candidate source window.
In the case the source window candidate is at the beginning of the source file, the extension of the search area to the,Ieft may not be possible and can be ignored. Similarly, in the case the source window candidate is at the end of the source file, the extension of the search area to the right may not be possible and can be ignored. FIG. 7 illustrates a continuation of the previous example used in describing method 200 to show the extended search area for the top 5 source window candidates and the associated target window T. 430 to 434 in FIG. 4 show the original top 5 source window candidates found by method 200. Now, by extending the search size by 2k to each candidate window, the extended search area for these top 5 source windc-w candidates 730 to 734 are shown in FIG. 7. For target window T, the search area for candidate source window 1 becomes So, Sl, S2, and S3. Note that the source window 1 search area cannot be extended to the left since it is already at the beginning of the file. Similarly, the search area for candidate source window becomes So, Sl, S2, S3, and S4; the search area for candidate source window 3 becomes S2, S3, S4, S5, and S6; the search area for candidate source window 4 becomes S6, S7, S8, S9, and Slo; the search area for candidate source window 5 becomes S8, S9, Sio, Sil, and S12.
In step 630, once the extended search area for each top source window candidate has been defined, method 500 can be used to find the best match between the extended search area for each of the top k candidates and the associated target window T. For candidate source window 1, the method uses method 500 to find a sub-segment with length t within the extended search area So, Sl, S2, and S3 that best matches the target window T.
Similarly, for candidate source window 2, the method uses method 500 to find a sub-segment with length t within the extended search area So, Sl, S2, S3, and S4 that best matches the target window T; for candidate source window 3, the method uses method 500 to find a sub-segment with lengt:h t within the Proprietary Information of AT&T

extended search area S2, S3, S4, S5, and S6 that best matches the target window T; for candidate source window 4, the method uses method 500 to find a sub-segment with length t within the extended search area S6, S7, S8, S9, and Slo that best matches the target window T; for candidate source window 5, the method uses method 500 to find a sub-segment with length t within the extended search area S8, S9, Sio, Si 1, and S12 that best matches the target window T. ;
In step 640, the sub-segment with the best distance produced in step 630 among all candidate extended search areas will be used as the best match to T. The method terminates in step 650.
FIG. 8 illustrates the pi-esent window matching method(s) implemented using a general purpose computer 800 or any other hardware equivalents.
For example, the present window matching methods and data structures can be represented by one or more software applications (or even a combination of software and hardware, e.g., using application specific integrated circuits (ASIC)), where the software is loaded from a storage medium 806, (e.g., a ROM, a magnetic or optical drive or diskette) and operated by the CPU 802 in the memory 804 of the system. As such, the present window matching methods and data structures of the present invention can be stored on a computer readable medium, e.g., RAM memory, ROM, magnetic: or optical drive or diskette and the like.
While various embodiments have been described above, it should be understood that they have been presented by way of example only, and not limitation. Thus, the breadth and scope of a preferred embodiment should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance with the following claims and their equivalents.
Proprietary Information of AT&T

Claims

1. A method for matching target file windows of a target file with source file windows of a source file in delta compressors, said method comprising:
partitioning the source file into a plurality of source file windows;
partitioning the target file into a plurality of target file windows; and matching one of said plurality of source file windows with one of said plurality of target file windows based upon a distance function that defines a distance between a data segment of said one of said plurality of source file windows and a data segment of said one of said plurality of target file windows using a respective n-gram of said data segments.

2. The method of claim 1, wherein said distance function is as follows:

where, .DELTA.(S1, S2) is defined to be the distance function between two non-empty data segments, S1 and S2, as a measure of similarity between the two data segments, and F S1 and F S2 are corresponding n-gram frequency arrays for data segments S1 and S2 respectively.

3. The method of claim 1, further comprising:
finding a minimum distance between a sub-segment of a source window S with length t that best matches a target window T based on the distance function.

4. The method of claim 1, further comprising:
performing a fast search for matching source windows.

5. The method of claim 4, wherein said fast search comprises a definition of a signature function:
where signature .sigma.(P) is defined as the sum of the hash values of all n-grams, F P is a frequency array of hash values indexed by n-grams, i represents the hashed value of n-grams and A is the size of the frequency array.

6. The method of claim 4, wherein said fast search comprises a definition of a signature distance function based on the signature function .sigma.:

where P1 and P2 represent two data segments.

7. The method of claim 4, wherein said fast search comprises the steps of:
a) partitioning source file S into a sequence of segments each of a fixed size .lambda.;
b) partitioning target window T into a sequence of segments each of a fixed size .lambda.;
c) computing the set of all S i's such that j <= i <= a-b and .delta.(S i, T i) <= .epsilon. and asserting that T j votes for each such position (i-j).lambda. as a candidate for a match for each segment T j in T; and d) taking the top k candidate positions and applying to a small neighborhood of them to find the best match to T.

8. The method of claim 7, wherein said parameter .epsilon. within said computing step is a graduated threshold value.

9. A computer-readable medium having stored thereon a plurality of instructions, the plurality of instructions including instructions which, when executed by a processor, cause the processor to perform the steps of a method for matching target file windows of a target file with source file windows of a source file in delta compressors, comprising of:
partitioning the source file into a plurality of source file windows;
partitioning the target file into a plurality of target file windows; and matching one of said plurality of source file windows with one of said plurality of target file windows based upon a distance function that defines a distance between a data segment of said one of said plurality of source file windows and a data segment of said one of said plurality of target file windows using a respective n-gram of said data segments.

10. The computer-readable medium of claim 9, wherein said distance function is as follows:

where, .DELTA.(S1, S2) is defined to be the distance function between two non-empty data segments, S1 and S2, as a measure of similarity between the two data segments, and F S1 and F S2 are corresponding n-gram frequency arrays for data segments S1 and S2 respectively.

11. The computer-readable medium of claim 9, further comprising:
finding a minimum distance between a sub-segment of a source window S with length t that best matches a target window T based on the distance function.

12. The computer-readable medium of claim 9, further comprising:
performing a fast search for matching source windows.

13. The computer-readable medium of claim 12, wherein said fast search comprises a definition of a signature function:

where signature .sigma.(P) is defined as the sum of the hash values of all n-grams, F P is a frequency array of hash values indexed by n-grams, i represents the hashed value of n-grams and A is the size of the frequency array.

14. The computer-readable medium of claim 12, wherein said fast search comprises a definition of a signature distance function based on the signature function .sigma.:

where P1 and P2 represent two data segments.

15. The computer-readable medium of claim 12, wherein said fast search comprises the steps of:
a) partitioning source file S into a sequence of segments each of a fixed size .lambda.;
b) partitioning target window T into a sequence of segments each of a fixed size .lambda.;
c) computing the set of all S i's such that j <= i <= a-b and .sigma.(S i, T
i) <= .epsilon. and asserting that T j votes for each such position (i-j).lambda.
as a candidate for a match for each segment T j in T; and d) taking the top k candidate positions and applying to a small neighborhood of them to find the best match to T.

16. The computer-readable medium of claim 15, wherein said parameter .epsilon.

within said computing step is a graduated threshold value.

17. An apparatus for matching target file windows of a target file with source file windows of a source file in delta compressors, comprising:
means for partitioning the source file into a plurality of source file windows;
means for partitioning the target file into a plurality of target file windows; and means for matching one of said plurality of source file windows with one of said plurality of target file windows based upon a distance function that defines a distance between a data segment of said one of said plurality of source file windows and a data segment of said one of said plurality of target file windows using a respective n-gram of said data segments.

18. The apparatus of claim 17, wherein said distance function is as follows:
where, .DELTA.(S1, S2) is defined to be the distance function between two non-empty data segments, S1 and S2, as a measure of similarity between the two data segments, and FS1 and FS2 are corresponding n-gram frequency arrays for data segments S1 and S2 respectively.

19. The apparatus of claim 17, further comprising:
means for finding a minimum distance between a sub-segment of a source window S with length t that best matches a target window T based on the distance function.

20. The apparatus of claim 17, further comprising:
means for performing a fast search for matching source windows.